lipilian / AnimeFaceGAN

Anime Face generation: Generates faces of Anime characters using Generative Adverserial Networks (GANs)

This Anime Face Generator will be trained based on DCGAN, visualize with visdom.

Data

The data used for generator training is collected by DCGAN_IMG_COLLECTOR.py, which collected raw anime image from konachan.

For face detection, DCGAN_FACE_DETECTOR.py was used for face feature detection, which thanks to the github User bchao1

Todo list

• DCGAN
• connect to visdom
• checkpoints saving
• Add Resnet structure to improve net
• Try condintional Gan net

Some resources

• self attention GAN in PyTorch
• gan hacks and tips
• batch norm
• CVPR 2018 Tutorial on GANs
• auxiliary classifier GAN ac gan from scratch with keras

Results

Progress during training (epochs 0 to 400):

96x96 Generated samples after 400 epochs. Training time: ~10 hours on one GTX 1080 GPUs